Media/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-09-15 13:01:45 +08:00
Code Issues Packages Projects Releases Wiki Activity

Choose cpu copy for 32 bit application using local memory.

Browse Source 
Change-Id: I74aed9475185b09d4569fafb3427052fff73fd89
Signed-off-by: Michal Mrozek <michal.mrozek@intel.com>
This commit is contained in:
Michal Mrozek
2020-02-24 13:56:41 +01:00
committed by sys_ocldev
parent 5751b5eb27
commit 0e85ccf084
2 changed files with 29 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
opencl/source/mem_obj/buffer.cpp
View File

@ -518,19 +518,22 @@ bool Buffer::isReadWriteOnCpuAllowed() {
}
bool Buffer::isReadWriteOnCpuPreffered(void *ptr, size_t size) {
//if buffer is not zero copy and pointer is aligned it will be more beneficial to do the transfer on GPU
if (!isMemObjZeroCopy() && (reinterpret_cast<uintptr_t>(ptr) & (MemoryConstants::cacheLineSize - 1)) == 0) {
return false;
}
if (MemoryPool::isSystemMemoryPool(graphicsAllocation->getMemoryPool())) {
//if buffer is not zero copy and pointer is aligned it will be more beneficial to do the transfer on GPU
if (!isMemObjZeroCopy() && (reinterpret_cast<uintptr_t>(ptr) & (MemoryConstants::cacheLineSize - 1)) == 0) {
return false;
}
//on low power devices larger transfers are better on the GPU
if (context->getDevice(0)->getDeviceInfo().platformLP && size > maxBufferSizeForReadWriteOnCpu) {
return false;
//on low power devices larger transfers are better on the GPU
if (context->getDevice(0)->getDeviceInfo().platformLP && size > maxBufferSizeForReadWriteOnCpu) {
return false;
}
return true;
}
//if we are not in System Memory Pool, it is more beneficial to do the transfer on GPU
//for 32 bit applications, utilize CPU transfers here.
if (!MemoryPool::isSystemMemoryPool(graphicsAllocation->getMemoryPool()) && is64bit) {
if (is64bit) {
return false;
}

18
opencl/test/unit_test/command_queue/read_write_buffer_cpu_copy.cpp
View File

@ -7,6 +7,7 @@
#include "shared/source/gmm_helper/gmm.h"
#include "shared/source/helpers/basic_math.h"
#include "shared/test/unit_test/helpers/debug_manager_state_restore.h"
#include "opencl/test/unit_test/command_queue/enqueue_read_buffer_fixture.h"
#include "opencl/test/unit_test/mocks/mock_command_queue.h"
@ -251,3 +252,20 @@ TEST(ReadWriteBufferOnCpu, given32BitApplicationWhenLocalMemoryPoolAllocationIsA
EXPECT_TRUE(buffer->isReadWriteOnCpuAllowed());
EXPECT_TRUE(buffer->isReadWriteOnCpuPreffered(reinterpret_cast<void *>(0x1000), MemoryConstants::pageSize));
}
TEST(ReadWriteBufferOnCpu, given32BitAppAndLocalMemoryBufferWhenAskedForCpuTransferPreferenceThenTrueIsReturned) {
if (is64bit) {
GTEST_SKIP();
}
DebugManagerStateRestore restorer;
DebugManager.flags.EnableLocalMemory.set(true);
auto device = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(nullptr));
MockContext ctx(device.get());
cl_int retVal = 0;
std::unique_ptr<Buffer> buffer(Buffer::create(&ctx, CL_MEM_READ_WRITE, MemoryConstants::pageSize, nullptr, retVal));
ASSERT_NE(nullptr, buffer.get());
EXPECT_TRUE(buffer->isReadWriteOnCpuAllowed());
EXPECT_TRUE(buffer->isReadWriteOnCpuPreffered(reinterpret_cast<void *>(0x1000), MemoryConstants::pageSize));
}